DECODING BRAIN SIGNALS FOR INNER SPEECH: EEG SIGNAL SUB-BAND DECOMPOSITION AND MACHINE LEARNING APPROACH

Decoding inner speech is a rapidly advancing field with significant implications for brain-computer interfaces (BCIs) and assistive technologies. In this study how sub-band decomposition, feature types, and classifier choice affect inner-speech classification were assessed. A comprehensive comparative framework is presented, employing various sub-band decomposition techniques including Tunable Q-Factor Wavelet Transform, Empirical Mode Decomposition, Variational Mode Decomposition, and Circulant Singular Spectrum Analysis. From the decomposed signals, 17 statistical, 4 entropy, and 5 frequency features were extracted individually and in hybrid combinations before reduction with Least Absolute Shrinkage and Selection Operator. Classification performance was evaluated with 10-fold cross-validation, using Support Vector Machine, Feedforward Neural Network, and K-Nearest Neighbor. Performance was summarized as the mean across 30 independent runs per condition over six metrics. A three-way repeated measures ANOVA with Greenhouse-Geisser correction showed significant main effects and interactions across all metrics (p < 0.001), with large effect sizes (η2p >= 0.14). The best overall configuration was EMD-Hybrid-SVM outperforming baseline models, yielding a group-level mean accuracy of 54.13%. These findings demonstrate the potential of sub-band decomposition and hybrid features for improving inner speech decoding and contribute to the development of more robust EEG-based BCIs.